3-Group Support: Select LoRAs from up to 3 different folders
Multi-Source Metadata Reading: Automatically retrieve trigger words and sample prompts with priority order:
1. .metadata.json format (ComfyUI Lora Manager) - Priority 1
2. .info format (Civitai Helper) - Priority 2
3. Embedded metadata in LoRA file - Priority 3 (NEW)
Strength Randomization: Randomize strength with range specification (e.g., 0.4-0.8)
Flexible Trigger Word Retrieval:
- json_combined: Combine all trigger word patterns (with deduplication)
- json_random: Randomly select one pattern
- json_sample_prompt: Randomly retrieve sample prompts
- metadata: Read directly from embedded metadata (NEW)
Sample Prompt Optimization: Automatically remove LoRA syntax from samples and apply node-configured strength
Dual Text Outputs: Separate positive_text and negative_text outputs
ComfyUI Standard Seed Control: Supports fixed/randomize/increment/decrement
Wildcard Encode Integration: Works seamlessly with Wildcard Encode (Inspire)
LoRA Syntax Auto-Removal: Cleans up LoRA syntax in additional_prompt to prevent noise

Installation

Requirements

ComfyUI (works with standard installation)
No additional libraries required ✅

Steps

cd ComfyUI/custom_nodes
git clone https://github.com/shin131002/RandomLoRALoader.git
# Or manually create RandomLoRALoader folder and copy files

Restart ComfyUI.

Note: Uses only Python standard library and ComfyUI bundled libraries, so no pip install needed.

Uninstallation

# Navigate to custom_nodes folder
cd ComfyUI/custom_nodes

# Remove RandomLoRALoader folder
rm -rf RandomLoRALoader

# On Windows
rmdir /s RandomLoRALoader

Or manually delete the RandomLoRALoader folder and restart ComfyUI.

Usage

Basic Usage

Add Node: Search for "Random LoRA Loader" in the node browser
Connect MODEL/CLIP: Connect base model and CLIP to inputs
Set Folder Paths:
- Group 1: Style LoRA folder path
- Group 2: Character LoRA folder path (optional)
- Group 3: Concept LoRA folder path (optional)
Configure Each Group:
- num_loras: Number of LoRAs to select from each group
- model_strength / clip_strength: Application strength (fixed value or random range)
Connect Outputs:
- MODEL / CLIP: Connect to KSampler
- positive / negative: Connect to Set Conditioning or KSampler
- positive_text / negative_text: Connect to Show Text for verification (recommended)

Wildcard Encode Integration (Recommended)

This node works seamlessly with Wildcard Encode (Inspire) for dynamic prompt generation combined with random LoRA selection.

Recommended Connection

[Wildcard Encode (Inspire)]
  text: "__style__, {red|blue|green}, <lora:base_effect:0.5>"
  ↓
  ├─ populated_text ──→ [Random LoRA Loader]
  │                     additional_prompt
  └─ MODEL/CLIP ──────→ model/clip input
  
[Random LoRA Loader]
  Folder 1: character LoRAs
  Folder 2: concept LoRAs
  ↓
  MODEL/CLIP/CONDITIONING → [KSampler]

How It Works

Wildcard Encode Processing:
- Wildcard expansion: __style__ → "anime style"
- Choice expansion: {red|blue|green} → "blue"
- LoRA syntax processing: <lora:base_effect:0.5> → applied to MODEL
- Result: populated_text = "anime style, blue, lora:base_effect:0.5"
Random LoRA Loader Processing:
- Auto-removes LoRA syntax from populated_text: "anime style, blue"
- Random LoRA selection: character_alice, concept_magic
- Final prompt: "anime style, blue, alice, blonde hair, magic circle"
- MODEL/CLIP: base_effect + character_alice + concept_magic (all applied)

Benefits

✅ Delegate wildcard functionality to Wildcard Encode (specialized tool)
✅ Delegate random LoRA selection to this node (simple configuration)
✅ Prompt information fully inherited
✅ LoRA syntax automatically cleaned (no noise)

3-Group Example

Group 1: style LoRAs
  - num_loras_1: 2
  - model_strength_1: "0.6-0.9"  ← Random
  - clip_strength_1: "0.6-0.9"   ← Random

Group 2: character LoRAs
  - num_loras_2: 1
  - model_strength_2: "1.0"      ← Fixed
  - clip_strength_2: "1.0"       ← Fixed

Group 3: Unused
  - lora_folder_path_3: (empty)
  - num_loras_3: 0

→ Result: 2 style LoRAs (variable strength) + 1 character LoRA (fixed strength) = 3 LoRAs total

Settings

Common Settings

Setting	Description	Default
`token_normalization`	Token normalization method	`none`
`weight_interpretation`	Prompt emphasis notation interpretation	`A1111`
`additional_prompt`	Additional prompt (combined with trigger words)	(empty)
`trigger_word_source`	Trigger word source	`json_combined`
`seed`	Random selection seed	`0`

Important Notes on additional_prompt

Supported:

Regular prompt text
Text input from other nodes (e.g., Wildcard Encode's populated_text)

Not Supported (automatically removed):

<lora:xxx:0.8> format LoRA syntax
{a|b|c} format wildcard syntax
__filename__ format wildcard syntax

Why:

This node applies LoRAs via folder specification
Wildcard functionality should be handled by dedicated nodes (Wildcard Encode, etc.)
Writing LoRA syntax causes filename to become meaningful tokens as noise

Example (problematic case):

Input: "1girl, <lora:anime_style:0.8>, beautiful"
↓
CLIP tokenization: [1girl, lora, anime, style, 0, 8, beautiful]
                            ^^^^^ ^^^^^
                            Unintended words added to prompt ❌

Correct usage:

additional_prompt: "1girl, beautiful"  ← No LoRA syntax ✅
LoRA application: Use folder specification feature ✅

Or, when connecting from Wildcard Encode, LoRA syntax is automatically removed:

[Wildcard Encode] populated_text: "1girl, <lora:style:0.8>, beautiful"
       ↓
[This Node] additional_prompt received → LoRA syntax auto-removed
       ↓
Final prompt: "1girl, beautiful" ✅

Group Settings (1-3)

Each group can be configured individually:

Setting	Description	Default
`lora_folder_path_X`	LoRA folder absolute path	(empty)
`include_subfolders_X`	Include subfolders	`true`
`model_strength_X`	MODEL application strength	`"1.0"`
`clip_strength_X`	CLIP application strength	`"1.0"`
`num_loras_X`	Number of LoRAs to select	Group 1: `1`, Groups 2/3: `0`

Strength Specification (Important)

Strength fields accept fixed values or random ranges.

Fixed Value

Input: "1.0"
→ Always applied at 1.0

Input: "0.55"
→ Always applied at 0.55

Features:

Input value used as-is
Up to 2 decimal places (e.g., 0.55)

Random Range

Input: "0.4-0.8"
→ Randomly selects from 0.4, 0.5, 0.6, 0.7, 0.8 (0.1 increments)

Input: "0.5-1.0"
→ Randomly selects from 0.5, 0.6, 0.7, 0.8, 0.9, 1.0

Input: "0.44-0.82"
→ Randomly selects from 0.4, 0.5, 0.6, 0.7, 0.8
   (Range automatically rounded to 1 decimal place)

Features:

Specify range with hyphen (-)
Values generated in 0.1 increments
Range bounds with 2+ decimal places are rounded to 1 decimal place
Different value selected each execution

Error Handling

Input: "abc" or invalid format
→ Error logged, processed with strength 1.0

Error message example:

[RandomLoRALoader] ❌ Cannot parse strength 'abc', using 1.0
[RandomLoRALoader] 💡 Usage: '1.0' or '0.4-0.8'

Verifying Actual Strength

Actual strength used can be verified in positive_text output:

<lora:style_anime:0.7:0.6>, anime style, vibrant,
               ↑   ↑
          MODEL    CLIP
          strength strength

Recommended: Connect Show Text node to positive_text to verify actual strength.

MODEL Strength vs CLIP Strength

Basic Concept

MODEL strength: Affects image itself (art style, composition, colors)
CLIP strength: Affects prompt understanding (trigger word effectiveness)

Recommended Settings

Beginners / General Use:

model_strength: "1.0"
clip_strength: "1.0"

Using same value for both is the standard approach.

Advanced / Fine-tuning:

model_strength: "0.7"
clip_strength: "1.0"

For cases like wanting subtle art style but strong trigger word effect.

Randomization for Experimentation:

model_strength: "0.6-0.9"
clip_strength: "0.6-0.9"

For testing various strength combinations.

Trigger Word Retrieval Methods

json_combined (All Combined)

Combines all trigger word patterns and removes duplicates.

Example:

"trainedWords": [
  "character, red hair, blue eyes, uniform",
  "character, red hair, blue eyes, casual clothes"
]

↓

Output: character, red hair, blue eyes, uniform, casual clothes

json_random (Random Selection)

Randomly selects one pattern from trigger words.

Example:

"trainedWords": [
  "character, red hair, blue eyes, uniform",
  "character, red hair, blue eyes, casual clothes"
]

↓

Output: character, red hair, blue eyes, casual clothes (random)

json_sample_prompt (Sample Retrieval)

Randomly retrieves one sample prompt from JSON.

Outputs both positive_text and negative_text
Automatically removes <lora:xxx:x.x> syntax from samples
Generates new LoRA syntax with node-configured strength

Example:

"images": [{
  "meta": {
    "prompt": "1girl, <lora:style:0.8>, beautiful",
    "negativePrompt": "bad quality, worst quality"
  }
}]

↓

positive_text: 1girl, beautiful (LoRA syntax removed)
negative_text: bad quality, worst quality

metadata (Embedded Metadata Only - NEW)

Reads trigger words directly from LoRA file's embedded metadata, ignoring external .metadata.json and .info files.

Supported metadata fields:

ss_tag_frequency: kohya_ss format tag frequency (extracts top 20 tags)
modelspec.trigger_word: Trigger word
ss_output_name: Model name

Use case:

When you want to use only the metadata embedded by the LoRA trainer
When external JSON files contain incorrect information
For LoRAs without external metadata files

Example:

LoRA file contains embedded metadata:
  ss_tag_frequency: {"dataset1": {"alice": 150, "blonde": 120, ...}}
  ↓
Output: alice, blonde, blue_eyes, ...

Note: Other modes (json_combined, json_random, json_sample_prompt) use external files with priority, falling back to embedded metadata if external files are not found.

Outputs

Output	Type	Description
`MODEL`	MODEL	Model with LoRA applied
`CLIP`	CLIP	CLIP with LoRA applied
`positive_text`	STRING	LoRA info and positive prompt
`negative_text`	STRING	Negative prompt (json_sample_prompt only)
`positive`	CONDITIONING	Positive conditioning
`negative`	CONDITIONING	Negative conditioning

positive_text Output Example

1girl, beautiful,
<lora:style_anime_v2:0.7:0.6>, anime style, vibrant colors,
<lora:character_alice:1.0:1.0>, alice, blonde hair, blue eyes,

Required File Structure

Metadata files or embedded data required for trigger word retrieval.

Priority Order

The node reads metadata in the following priority order:

.metadata.json (ComfyUI Lora Manager format)
.info (Civitai Helper format)
Embedded metadata in LoRA file (Always available if LoRA was trained with metadata)

1. ComfyUI Lora Manager format (Recommended)

/path/to/lora/
├── style_anime_v2.safetensors
├── style_anime_v2.metadata.json  ← ComfyUI Lora Manager
├── character_alice.safetensors
└── character_alice.metadata.json

2. Civitai Helper format

/path/to/lora/
├── style_anime_v2.safetensors
├── style_anime_v2.info  ← Civitai Helper
├── character_alice.safetensors
└── character_alice.info

3. Embedded metadata (No external file needed)

/path/to/lora/
├── style_anime_v2.safetensors  ← Contains embedded metadata
└── character_alice.safetensors  ← Contains embedded metadata

Note:

If external files (.metadata.json or .info) exist, they take priority over embedded metadata (except when using trigger_word_source: metadata)
When trigger_word_source is set to metadata, only embedded metadata is read, ignoring external files

Generating metadata files

ComfyUI Lora Manager (Recommended):

ComfyUI Lora Manager
Generates .metadata.json files

Civitai Helper:

Civitai Helper
Generates .info files (also supported)

Embedded metadata:

Automatically created during LoRA training with kohya_ss or other trainers
No additional tools required

Troubleshooting

Cannot retrieve trigger words

For external metadata files:

Check if metadata file exists (.metadata.json or .info)
Verify JSON contains civitai.trainedWords or civitai.images
Confirm filename matches {LoRAfilename}.metadata.json or {LoRAfilename}.info

For embedded metadata:

Set trigger_word_source to metadata to read directly from LoRA file
Check if LoRA was trained with metadata (most modern LoRAs include this)
Supported fields: ss_tag_frequency, modelspec.trigger_word, ss_output_name

additional_prompt with LoRA syntax doesn't work

Symptom:

additional_prompt: "<lora:anime_style:0.8>, 1girl"
→ anime_style not applied

Cause: This node doesn't interpret <lora:xxx:0.8> syntax. LoRA syntax is automatically removed.

Solution:

Use folder specification feature for LoRA application
When integrating with Wildcard Encode (Inspire), let it handle LoRA syntax

Unintended words in prompt

Symptom:

additional_prompt: "<lora:anime_style:0.8>"
→ "anime", "style" words automatically added

Cause: v1.0.0 and later automatically remove LoRA syntax, so this issue doesn't occur. If using older version, LoRA syntax filename part was interpreted as tokens.

Solution:

Update node to latest version
Don't write LoRA syntax in additional_prompt

LoRA not applied

Verify folder path is correct
Check num_loras isn't 0
Check ComfyUI console for error logs

Strength not as intended

Check positive_text output with Show Text
Displays actual strength like <lora:xxx:0.7:0.6>
With random range specification, different value selected each time
Check console for strength selection log: [RandomLoRALoader] Selected 0.6 from range 0.4-0.8

Strength input error

Symptom:

[RandomLoRALoader] ❌ Cannot parse strength 'abc', using 1.0

Solution:

Use correct format: "1.0" or "0.4-0.8"
Range specification must use hyphen (-)
Node won't crash with invalid input, operates with 1.0

Default values not displayed

Completely restart ComfyUI
Clear browser cache (Ctrl+Shift+R)
Delete and re-add node

Console shows many "lora key not loaded" messages

Symptom:

lora key not loaded: lora_unet_input_blocks_4_1_proj_in.alpha
lora key not loaded: lora_unet_input_blocks_4_1_proj_in.lora_down.weight
...

Cause and Solution:

These messages are compatibility warnings output by ComfyUI core and don't affect functionality. Displayed when some keys in LoRA file aren't compatible with current model, but compatible keys apply normally.

v1.0.0 and later automatically suppress these warning messages. If still displayed in large quantities:

Verify node file is latest version
Restart ComfyUI
If still appears and functionality is fine, safe to ignore

Tips & Tricks

Random LoRA Loader: Using as a LoRA Syntax Filter Only

You can use this node as a simple LoRA syntax remover without applying any LoRAs.

Setup:

Set all num_loras to 0 (or leave all folder paths empty)
Connect Wildcard Encode's populated_text to additional_prompt
Connect positive_text output to next node

Use case:

Remove LoRA syntax from Wildcard Encode output
Clean up prompts before passing to other nodes
Use Wildcard Encode's LoRA processing while removing the syntax for compatibility

Example workflow:

[Wildcard Encode (Inspire)]
  text: "__style__, {red|blue|green}, <lora:base_effect:0.5>"
  ↓
  populated_text: "anime style, blue, <lora:base_effect:0.5>"
  (LoRA already applied to MODEL by Wildcard Encode)
  ↓
[Random LoRA Loader]
  num_loras: 0, 0, 0  ← No LoRA application
  additional_prompt ← Connected from populated_text
  ↓
  positive_text: "anime style, blue,"  ← LoRA syntax removed ✅
  ↓
[Next Node]

Benefits:

✅ Removes <lora:filename:strength> syntax that can become noise tokens
✅ No need for separate text processing nodes
✅ Works with any text containing LoRA syntax

Filtered Random LoRA Loader (NEW v1.1.0)

Overview

A single-group LoRA loader with keyword filtering capabilities. Designed for chaining multiple instances together.

Key differences from Random LoRA Loader:

Single folder path (not 3)
Keyword filtering with AND/OR modes
Metadata search capability with caching
Dual prompt inputs (positive and negative)
Recommended for serial connection

Parameters

Common Settings

token_normalization - Same as Random LoRA Loader
weight_interpretation - Same as Random LoRA Loader

Folder Settings

lora_folder_path - Single folder path
include_subfolders - Include subfolders (default: True)

Keyword Filter

keyword_filter - Keywords (comma-separated, e.g., "style, anime")
filter_mode - AND (all keywords) or OR (any keyword, default: AND)
search_in_metadata - Search in metadata files (default: False - filename only)

LoRA Settings

model_strength - MODEL strength (e.g., "1.0" or "0.6-0.9")
clip_strength - CLIP strength (e.g., "1.0" or "0.6-0.9")
num_loras - Number of LoRAs to select (default: 1, range: 0-20)

Trigger Word Settings

trigger_word_source - Same as Random LoRA Loader

Additional Prompts

additional_prompt_positive - Additional positive prompt (input or direct entry)
additional_prompt_negative - Additional negative prompt (input or direct entry)

Seed

seed - Random seed
control_after_generate - Seed control mode

Keyword Filtering

Filename Search (Default - Fast)

Searches only in LoRA filename.

Example:

Folder contents:
  - style_anime_v1.safetensors
  - style_realistic.safetensors
  - character_alice.safetensors

keyword_filter: "anime"
→ Matches: style_anime_v1.safetensors ✅

Metadata Search (Optional - Slower)

Searches in both filename and metadata fields:

model_name
civitai.name
civitai.trainedWords
civitai.model.name
civitai.model.tags
tags
Embedded metadata

Performance Notes:

Initial loading time (SSD environment, with progress display):

100 files: Instant
1,000 files: ~2 seconds
5,000 files: ~10 seconds
10,000 files: ~20 seconds
20,000 files: ~40 seconds
50,000 files: ~100 seconds (1 min 40 sec)

Memory usage:

10,000 files: ~150MB
20,000 files: ~300MB
50,000 files: ~750MB

Second time onward: Instant (cached)

Recommendations:

Default filename search is always fast
Enable metadata search only when needed
For large collections (5,000+ files), initial loading will take time
HDD environments take approximately 3x longer than SSD for initial loading

Example:

File: abc123.safetensors
Metadata: {"trainedWords": ["anime style", "vibrant colors"]}

search_in_metadata: True
keyword_filter: "anime"
→ Matches ✅

AND vs OR Modes

AND Mode (default):

keyword_filter: "style, anime"
filter_mode: "AND"
→ Matches files containing BOTH "style" AND "anime"

OR Mode:

keyword_filter: "style, anime"
filter_mode: "OR"
→ Matches files containing EITHER "style" OR "anime"

Metadata Search Performance

Filename search (Default):

search_in_metadata: False (default)
Very fast even with 10,000+ LoRAs
Processing time: <100ms

Metadata search:

search_in_metadata: True
Searches in both filename and LoRA metadata
First execution: Builds cache
- 1,000 LoRAs: ~2 seconds
- 10,000 LoRAs: ~7-75 seconds (depends on metadata files)
- Progress displayed in console
Subsequent executions: Very fast (<100ms)
Cache persists until ComfyUI restart

Performance tip:

Use include_subfolders: True on first run to cache all LoRAs at once
After cache is built, include_subfolders: False will still use cached metadata
Cache is shared across all node instances

Usage Examples

Basic Usage

[Filtered Random LoRA Loader]
  lora_folder_path: /loras/all/
  keyword_filter: "anime"
  num_loras: 1
  ↓
Randomly selects 1 LoRA containing "anime" in filename

Chaining Multiple Instances

[Filtered Random LoRA Loader A]
  folder: /loras/all/
  keyword_filter: "style, anime"
  filter_mode: "AND"
  num_loras: 1
  ↓
[Filtered Random LoRA Loader B]
  folder: /loras/all/
  keyword_filter: "character"
  num_loras: 1
  additional_prompt_positive ← A's positive_text
  additional_prompt_negative ← A's negative_text
  ↓
[Filtered Random LoRA Loader C]
  folder: /loras/all/
  keyword_filter: "concept"
  num_loras: 1
  additional_prompt_positive ← B's positive_text
  additional_prompt_negative ← B's negative_text

With Wildcard Encode

[Wildcard Encode]
  populated_text: "1girl, beautiful"
  negative_text: "bad quality"
  ↓
[Filtered Random LoRA Loader]
  additional_prompt_positive ← populated_text
  additional_prompt_negative ← negative_text
  keyword_filter: "style"
  ↓
Combined output ✅

Best Practices

Use descriptive filenames: style_anime_v1.safetensors
Default mode (fast): Keep search_in_metadata: False for daily use
Enable metadata search when:
- Filenames don't contain useful keywords
- Need to search in trigger words or tags
First run optimization:
- Use include_subfolders: True once to build complete cache
- Switch to False after for faster folder scanning
Chaining:
- Connect positive_text → next node's additional_prompt_positive
- Connect negative_text → next node's additional_prompt_negative

Troubleshooting

No LoRAs found matching filter:

Check if keyword exists in filenames
Enable search_in_metadata: True to search in metadata
Try OR mode instead of AND mode
Verify folder path is correct

First execution is slow:

This is normal when search_in_metadata: True
Cache is being built, subsequent executions will be fast
Progress is shown in console

Cache not working:

Cache clears on ComfyUI restart
Run once after startup to rebuild cache

Tips & Tricks (Original Node)

Using as a LoRA Syntax Filter Only

You can use this node as a simple LoRA syntax remover without applying any LoRAs.

Setup:

Set all num_loras to 0 (or leave all folder paths empty)
Connect Wildcard Encode's populated_text to additional_prompt
Connect positive_text output to next node

Use case:

Remove LoRA syntax from Wildcard Encode output
Clean up prompts before passing to other nodes
Use Wildcard Encode's LoRA processing while removing the syntax for compatibility

Example workflow:

[Wildcard Encode (Inspire)]
  text: "__style__, {red|blue|green}, <lora:base_effect:0.5>"
  ↓
  populated_text: "anime style, blue, <lora:base_effect:0.5>"
  (LoRA already applied to MODEL by Wildcard Encode)
  ↓
[Random LoRA Loader]
  num_loras: 0, 0, 0  ← No LoRA application
  additional_prompt ← Connected from populated_text
  ↓
  positive_text: "anime style, blue,"  ← LoRA syntax removed ✅
  ↓
[Next Node]

Benefits:

✅ Removes <lora:filename:strength> syntax that can become noise tokens
✅ No need for separate text processing nodes
✅ Works with any text containing LoRA syntax

Disclaimer and Support Policy

About This Node

This custom node was developed for personal use and is being shared publicly.

Support

❌ No technical support provided
- Usage questions
- Environment setup assistance
- Individual troubleshooting support
❌ No warranty
- Operation in specific environments
- Compatibility with future ComfyUI updates
- Liability for data loss, etc.

Community Contributions

The following are welcome but not guaranteed to be addressed:

✅ Bug reports (Issues)
✅ Pull requests
✅ Feature suggestions

Terms of Use

Use entirely at your own risk
Test thoroughly before production use
Author assumes no liability for any issues

License

MIT License

This software is provided "as is" without any express or implied warranties.

Changelog

v1.0.0 (2024-12-30)

Initial public release
3-group support (select from different folders)
Strength randomization (range specification: 0.4-0.8)
External JSON reading (Civitai Helper compatible)
Auto-remove LoRA syntax in sample prompts
Auto-remove LoRA syntax in additional_prompt (noise prevention)
Separate positive_text/negative_text outputs
Wildcard Encode (Inspire) integration
Fully local operation (no API required)

References

ComfyUI
ComfyUI Lora Manager - Generates .metadata.json (Recommended)
Civitai Helper - Generates .info (Also supported)